Maximizing Hamming Distance in Contraction of

Permutation Arrays
∗ † ∗ ∗
Sergey Bereg Zevi Miller Luis Gerardo Mojica Linda Morales
I.H. Sudborough∗
April 19, 2018
arXiv:1804.03768v2 [math.CO] 19 Apr 2018

Abstract
Let M (n, d) be the maximum size of a permutation array on n symbols with pairwise
Hamming distance at least d. In this paper we study permutation arrays produced by
applying the contraction operation [2] to the groups AGL(1, q) and P GL(2, q) for a prime
power q satisfying q ≡ 1 (mod 3). We introduce contraction graphs and show that they
have large independent sets. We obtain the following lower bounds
1. M (q − 1, q − 3) ≥ (q 2 − 1)/2 for odd q ≥ 7,
2. M (q − 1, q − 3) ≥ (q − 1)(q + 2)/3 for q = 22k ,
3. M (q, q − 3) ≥ Kq 2 log q for some constant K if q ≥ 13 is odd.
We also obtain lower bounds on M (n, d) for some pairs (n, d) by applying the contraction
operation to the Mathieu groups.

1 Introduction
1.1 Notation and General Background
We consider permutations on a set Ω of size n. Given two such permutations π and σ, we let
hd(π, σ) = |{x ∈ Ω : π(x) 6= σ(x)}|, so hd(π, σ) is the number of elements of Ω at which π and
σ disagree. When hd(π, σ) = d, we say that π and σ and are at Hamming distance d, or that
the Hamming distance between π and σ is d. A permutation array A is set of permutations on
Ω. We say that hd(A) = d if d = min{hd(π, σ) : π, σ ∈ A}. For positive integers n and d with
d ≤ n we let M (n, d) be the maximum number of permutations in any permutation array A
satisfying hd(A) ≥ d.
Consider a fixed ordering x1 , x2 , · · · , xn of the elements of Ω. The image string of the per-
mutation σ ∈ A is the string σ(x1 )σ(x2 ) · · · σ(xn ). Thus the permutation array A can also be
regarded as an |A| × n matrix whose rows are the image strings of the permutations in A. When
hd(A) = d, any two rows of A disagree in at least d positions and some pair of rows disagree in
exactly d positions.
The study of permutation arrays began (to our knowledge) with the papers [9] and [13],
where good bounds on M (n, d) (together with other results) were developed based on combina-
torial methods, motivated by the Gilbert-Varshamov bounds for binary codes. In recent years
∗
Computer Science Dept., University of Texas at Dallas, Richardson, TX 75083, USA
†
Dept. of Mathematics, Miami University, Oxford, OH 45056, USA

1
there has been renewed interest in permutation arrays, motivated by applications in power line
transmission [12], [20], [27], and [15], block ciphers [26], and in multilevel flash memories [17]
and [18].
We review here some of the known results and methods for estimating M (n, d).
Some elementary exact values and bounds on M (n, d) are the following (summarized with
short proofs in [8]) ; M (n, 2) = n!, M (n, 3) = n!2 , M (n, n) = n, M (n, d) ≥ M (n − 1, d),
n!
M (n, d) ≥ M (n, d + 1), M (n, d) ≤ nM (n − 1, d), and M (n, d) ≤ (d−1)! . More sophisticated
bounds were developed in the above cited papers [9] and [13], with a recent improvement in [28].
The smallest interesting case for d is d = 4. Here some interesting and non-elementary bounds
for M (n, 4) were developed in [11], using linear programming, characters on the symmetric group
2
Sn , and Young diagrams. In [19] it√is shown that if K > 0 is a constant with n > e30/K and
s < n1−K , then M (n, n − s) ≥ θ(s! log n). The lower bound is achieved by a polynomial time
randomized construction, using the Lovasz Local Lemma in the analysis.
There are various construction methods for permutation arrays. First there is a connection
with mutually orthogonal latin squares (MOLS). It was shown in [7] that if there are m MOLS
of order n, then M (n, n − 1) ≥ mn. From this it follows that if q is a prime power, then
M (q, q − 1) = q(q − 1). Computational approaches for bounding M (n, d) for small n and d,
including clique search, and the use of automorphisms are described in [8], [16], and [22]. There
are also constructions of permutation arrays that arise from the use of permutation polynomials,
also surveyed in [8], which we mention briefly below.
Additional construction methods are coset search [2] and partition and extension [3]. In the
first of these, one starts with with a permutation group G on n letters with hd(G) = d, and
which is a subgroup of some group H (for example H = Sn ). Now for disjoint permutation
arrays A, B on the same set of letters, let hd(A, B) = min{hd(σ, τ ) : σ ∈ A, τ ∈ B}. For x ∈ /G
we observe that the coset xG of G in H is a permutation array with hd(xG) = hd(G). For
cosets x1 G, x1 G, · · · , xk G of G, the Hamming distance of the permutation array ∪1≤i≤k xi G is
the minimum of d and m, where m = min{hd(xi G, xj G) : 1 ≤ i < j ≤ k}. The method of coset
search is to iteratively find coset representatives xi so that m, while in general less than d, is
still reasonably large. The partition and extension method is a way of obtaining constructive
lower bounds M (n + 1, d + 1) from such bounds for M (n, d).
Moving closer to the subject of this paper, we consider a class of optimal constructions which
arise through sharply transitive groups. We say that a permutation group G acting on a set Ω
is sharply k-transitive on Ω if given any two k-tuples x1 , x2 , · · · , xk and y1 , y2 , · · · , yk of distinct
elements of Ω there exists a unique g ∈ G such that g(xi ) = yi for all 1 ≤ i ≤ k. The following
theorem will be used in this paper.

Theorem 1 [9] Suppose G is a permutation group which is sharply k-transitive on a set of size
n, n ≥ k + 2. Then hd(G) = n − k + 1 and M (n, n − k + 1) = |G|.

The set of all sharply k-transitive groups (for k ≥ 2) are known, and are as follows [6], [10];
k = 2 : the Affine General Linear Group AGL(1, q) acting on the finite field GF (q), consisting
of the transformations {x → ax + b : x, a 6= 0, b ∈ GF (q)},
k = 3 : the Projective Linear Group P GL(2, q) acting on GF (q) ∪ {∞}, consisting of the
transformations {x → ax+b
cx+d
: x, a, b, c, d ∈ GF (q), ad − bc 6= 0},
k = 4 : the Mathieu group M11 acting on a set of size 11,
k = 5 : the Mathieu group M12 acting on a set of size 12,
arbitrary k : the symmetric group Sk acting on a set of size k is sharply k and (k − 1)-transitive,
as well as the alternating group Ak acting on a set of size k is sharply (k − 2)-transitive.

2
 In this paper we obtain new lower bounds on M (n, d) for n and d near a prime power.
Previous results of this kind are given in [8] where it is shown that for n = 2k with n 6≡ 1(mod
3) we have M (n, n − 3) ≥ (n + 2)n(n − 1) and M (n, n − 4) ≥ 13 n(n − 1)(n2 + 3n + 8). It is also
shown that for any prime power n with n 6≡ 2(mod 3) we have M (n, n − 2) ≥ n2 . These results
are based on permutation polynomials.
Several results are shown in [2] using a contraction operation to AGL(1, q) and P GL(2, q).
1. M (n − 1, n − 3) ≥ n(n − 1) for a prime power n 6≡ 1 (mod 3),
2. M (n − 2, n − 5) ≥ n(n − 1) for a prime power n such that n ≡ 2 (mod 3) and n 6≡ 0, 1
(mod 5), and
3. M (n, n − 3) ≥ (n + 1)n(n − 1) for a prime power n 6≡ 1 (mod 3).
Our method is to apply a contraction operation to the groups AGL(1, q) and P GL(2, q). We
obtain the following lower bounds, assuming that q is prime power satisfying q ≡ 1 (mod 3);
1. for q ≥ 7, M (q − 1, q − 3) ≥ (q 2 − 1)/2 for q odd and M (q − 1, q − 3) ≥ (q − 1)(q + 2)/3
for q even, and
2. for q ≥ 13, M (q, q − 3) ≥ Kq 2 log q for some constant K if q is odd, and
3. bounds for M (n, d) for a finite number of exceptional pairs n, d, obtained from the Mathieu
groups.

1.2 Contraction
Consider a permutation array A acting on a set Ω = {x1 , x2 , · · · , xn } of size n, where the elements
of Ω are ordered by their subscripts. We distinguish some element, say xn , by renaming it F .
Thus the image string of any element σ ∈ A will be σ(x1 )σ(x2 ) · · · σ(F ), and we say that σ(xi )
occurs in position or coordinate xi of the string. Now for any π ∈ A, define the permutation π 4
on Ω by


π(F )
 if π(x) = F,
4
π (x) = F if x = F,

π(x) otherwise.


Thus the image string of π 4 is obtained from the image string of π by interchanging the
symbols F and π(F ) if π(F ) 6= F , while π 4 = π if and only if π(F ) = F . In either case, π 4
4
has F as its final symbol. We let π− be the permutation on n − 1 symbols obtained from π 4
by dropping the last symbol F from π 4 . As an example, if π = aF bcd, then π 4 = adbcF , and
4 4 4
π− = adbc. Further, for any subset R ⊂ A, let R4 = {π 4 : π ∈ R}, and R− = {π− : π ∈ R}.
4
So R− is a permutation array on the n − 1 symbols acting on the set Ω − {F }, and is called the
contraction of R.
We note some basic properties of the contraction operation.
Lemma 2 Let G be a permutation group acting on the set Ω of size n, and π, σ ∈ G.
a) The only coordinates in either π or σ whose values are affected by the 4 operation are
π −1 (∞), σ −1 (∞), and ∞. Thus hd(π 4 , σ 4 ) ≥ hd(π, σ) − 3.
b) Assume hd(π 4 , σ 4 ) = hd(π, σ) − 3. Then πσ −1 contains a 3-cycle in its disjoint cycle
factorization, and |G| is divisible by 3.
c) Let S ⊆ G. Then |S 4 | = |S−4 | and hd(S 4 ) = hd(S−4 ). If also hd(S) > 3, then |S| = |S 4 |.

3
Proof. Part a) follows immediately from the definition of the 4 operation.
For b), the assumption implies that there are positions xi , xj , F at which the image strings
of π and σ disagree and π 4 and σ 4 agree. So for some indices s, t we must have π(xi ) =
xs , π(xj ) = F, π(F ) = xt , while σ(xi ) = F, σ(xj ) = xt , σ(F ) = xs . Then πσ −1 (composing left
to right) contains the 3-cycle (xi , F, xj ) in its disjoint cycle factorization. Thus the subgroup
of G generated by πσ −1 has order divisible by 3, and hence |G| is divisible by 3 by Lagrange’s
theorem.
Consider c). The first two equalities follow from the fact that all image strings in S 4 have
F as their last coordinate. To see |S| = |S 4 | when hd(S) > 3, suppose to the contrary that
π 4 = σ 4 for distinct π, σ ∈ S. As noted in the proof of part a), π 4 and σ 4 can agree in at
most three positions where π and σ disagreed. Thus π and σ already agreed in at least n − 3
positions. So hd(π, σ) ≤ 3, a contradiction.

2 The contraction graph for AGL(1, q)

Recall the Affine General Linear Group AGL(1, q) acting as permutation group on the finite
field GF (q) of size q, as the set of transformations {x → ax + b : x, a 6= 0, b ∈ GF (q)} under
the binary operation of composition. Clearly |AGL(1, q)| = q(q − 1), and it is easy to show that
AGL(1, q) is sharply 2-transitive in this action. It follows from Theorem 1 (and is easy to verify
directly) that hd(AGL(1, q)) = q − 1. For any π ∈ AGL(1, q) the permutation π 4 on GF (q) is
defined as in the introduction, based on some ordering x1 , x2 , · · · , xq of the elements of GF (q),
where F = xq is a distinguished element.
Our goal in this section is to obtain a lower bound on M (q − 1, q − 3) for prime powers q ≥ 7
satisfying q ≡ 1(mod 3). Our method will involve the contraction of AGL(1, q). For brevity set
H = AGL(1, q).
We begin with some motivation for the assumption q ≡ 1(mod 3) in the results of this section.
We claim that if q 6≡ 1(mod 3), then hd(H 4 ) ≥ q − 3. If not, using hd(H) = q − 1 we have by
Lemma 2a that hd(H 4 ) = q − 4, and for some π, σ ∈ H we have hd(π 4 , σ 4 ) = hd(π, σ) − 3,
and hd(π, σ) = 1. Let x be the coordinate of agreement between π and σ, and let Hx be the
stabilizer subgroup for x; that is, the subgroup of H fixing x. By the sharp 2-transitivity of H
we have |Hx | = q − 1, and by Lemma 2, πσ −1 contains a 3-cycle in its cycle factorization. Since
πσ −1 ∈ Hx it follows by Lagrange’s theorem that 3 divides |Hx | = q − 1, a contradiction.
So with q 6≡ 1(mod 3) we have that H−4 is a permutation array on q − 1 symbols satisfying
hd(H−4 ) = hd(H 4 ) ≥ q − 3 by Lemma 2c and the claim of the preceding paragraph. Also by
Lemma 2c we have |H−4 | = |H 4 | = |H| = q(q − 1) since q ≥ 7. So trivially we get the bound
M (q − 1, q − 3) ≥ |H−4 | = q(q − 1).
So in this paper we investigate the case where q ≡ 1(mod 3), where this simple lower bound
for M (q − 1, q − 3) is not available.
The key idea is to find a subset I ⊂ AGL(1, q) such that hd(I 4 ) ≥ q − 3. Once we have such
an I, then I−4 is a permutation array on q − 1 symbols, and by Lemma 2c satisfies hd(I−4 ) =
hd(I 4 ) ≥ q − 3. This implies the lower bound M (q − 1, q − 3) ≥ |I−4 | = |I 4 | = |I|, the last
equality by Lemma 2c, since q ≥ 7 implies hd(I) ≥ q − 1 > 3. The actual size of I will then
yield our precise lower bound.
We will find I as an independent set in the following graph. Consider the contraction graph
CA (q) defined by V (CA (q)) = AGL(1, q), and E(CA (q)) = {πσ : hd(π 4 , σ 4 ) = q−4}. Now recall
that hd(AGL(1, q)) = q−1, so by Lemma 2a we have hd(π 4 , σ 4 ) ≥ q−4 for all π, σ ∈ AGL(1, q).
So if I is an independent set in CA (q), then hd(I 4 ) ≥ q − 3.

4
 We are thus reduced to finding a large independent set I in CA (q), and from this we get
the bound M (q − 1, q − 3) ≥ |I|. We begin on that in the following Lemma, which establishes
relations in the the finite field GF (q) that correspond to edges in the graph CA (q).

Lemma 3 Let π and σ be vertices of the graph CA (q), q ≡ 1(mod 3), say with σ(x) = ax + r
and π(x) = bx + s.
a) If a 6= b, then hd(π, σ) = q − 1.
b) If π(F ) = F , then π is an isolated point in CA (q). There are at least q − 1 points π satisfying
π(F ) = F .
c) Suppose π and σ are neighbors in CA (q). Then
c1) hd(π, σ) = q − 1, and hd(π 4 , σ 4 ) = hd(π, σ) − 3, and
c2) ab and ab are the distinct roots of the quadratic t2 + t + 1 = 0 over GF (q).
s−r
Proof. For a), just observe that π(x) = σ(x) has the unique solution x = a−b .
For the first claim in b), suppose not. Then for some vertex σ we have hd(π 4 , σ 4 ) = q − 4,
implying also that hd(π, σ) = q − 1 by Lemma 2a. Let i be the coordinate of agreement
between π and σ. Since π(F ) = F , we have π 4 = π. Thus hd(π, σ 4 ) = q − 4. Now σ 4 can
have only at most two coordinates, apart from i, in which it agrees with π, these being F and
j, where σ(j) = F . So altogether π and σ 4 agree in at most the 3 coordinates i, j, F . So
q − 4 = hd(π, σ 4 ) ≥ q − 3, a contradiction.
Now consider the second claim in b). Given that π(F ) = F , we have q − 1 choices for the
value π(i) for any fixed i ∈ GF (q), i 6= F . For each such choice the permutation π is uniquely
determined by the sharp 2-transitivity of AGL(1, q) acting on GF (q). The claim follows.
For c1), by the definition of edges in CA (q) we have q − 4 = hd(π 4 , σ 4 ) ≥ hd(π, σ) − 3 using
Lemma 2a. Since hd(π, σ) = q or q − 1, it follows that hd(π, σ) = q − 1 and we have equality
throughout, as required.
Consider c2). By part c1) we have hd(π, σ) = q − 1 and hd(π 4 , σ 4 ) = hd(π, σ) − 3. So there
are distinct α, β ∈ GF (q), with neither α nor β being F , such that σ(F ) = i, σ(α) = F , and
σ(β) = j, and π(F ) = j, π(α) = i, and π(β) = F for distinct i, j ∈ GF (q). This gives the
following set of equations in GF (q).


 σ(α) − σ(β) = F − j = a(α − β)
σ(α) − σ(F ) = F − i = a(α − F )

(1)

 π(α) − π(β) = i − F = b(α − β)
π(α) − π(F ) = i − j = b(α − F ).


The second and third equations of (1) imply

a(α − F ) = −b(α − β). (2)

Now starting with the first equation of (1) we obtain

a(α − β) = F − j
= (F − i) + (i − j)
= (a + b)(α − F ) (by the second and fourth equations of (1)).

Multiplying equation (2) by a and the last equation by b, we obtain the equations
 2
a (α − F ) = −ab(α − β)
(3)
ab(α − β) = b(a + b)(α − F ).

5
 Thus a2 (α − F ) = −b(a + b)(α − F ), and on dividing by α − F (since α 6= F ) we obtain

a2 + b(a + b) = 0. (4)

Dividing equation (4) by a2 or by b2 , we obtain that a/b and b/a are both roots of the equation
t2 + t + 1 = 0.
To see that a/b and b/a are distinct, it suffices to observe that 1 cannot be a root of t2 +t+1 = 0
under our assumptions. Assuming otherwise, if q is even then from 1 + t + t2 = 0 we get the
contradiction 1 = 0 since the characteristic is 2. If q is odd, then√by the assumption q ≡ 1(mod
3) and Corollary 19a the roots of t2 + t + 1 = 0 are t = 12 (−1 ± −3). Neither of these roots is
1, again since q is odd and q ≡ 1(mod 3).

We now define a graph H(q) on the same vertex set as that of CA (q) which contains CA (q) as
a subgraph. Then any independent set in H(q) is also independent in CA (q), and thus it suffices
to find a suitably large independent set in H(q). By Corollary 19a, the equation t2 + t + 1 = 0
has two distinct roots in GF (q) for q ≡ 1 (mod 3). Also direct substitution shows that if t is a
root of this equation, then so is 1t . So with Lemma 3 as motivation, let us call two permutations
π, σ ∈ AGL(1, q), say with σ(x) = ax + r and π(x) = bx + s, associates if ab and ab are the
distinct roots of the quadratic t2 + t + 1 = 0 over GF (q). We then let V (H(q)) = AGL(1, q)
and E(H(q)) = {πσ : π and σ are associates}. By Lemma 3 we know that CA (q) is a subgraph
of H(q).

Theorem 4 Let q be a prime power with q ≡ 1 (mod 3). Then

a) H(q) is regular of degree 2.
b) Every connected component of H(q) is a cycle C3s of length a multiple of 3.
c) Let q be odd, and D a connected component of H(q). If D is also a connected component of
CA (q), then D ∼
= C6r for some integer r ≥ 0.

Proof. For part a), let π ∈ AGL(1, q) be given by π(x) = ax + r, and let t1 and t2 be the
distinct roots of t2 + t + 1 = 0 over GF (q). Then by the definition of edges in H(q) we know that
the neighbors of π in H(q) are of the form σ1 (x) = at1 x + r1 and σ2 (x) = at2 x + r2 = ta1 x + r2 ,
so far with arbitrary r1 , r2 ∈ GF (q). By Lemma 3a we have hd(π, σ1 ) = hd(π, σ2 ) = q − 1. Let
u and v be such that π(u) = σ1 (u) and π(v) = σ2 (v). Then r1 and r2 are determined uniquely
by r1 = au(1 − t1 ) + r and r2 = av(1 − t11 ) + r. It follows that π has degree 2 as required.
Now consider b). Let D be a connected component of H(q). By part a) D must be a cycle,
and it remains to prove the claim about the length of D. Clearly |V (D)| ≥ 3 since H(q) is a
simple graph. We can suppose that |V (D)| ≥ 4 since otherwise |V (D)| = 3 and there is nothing
to prove.
Now take an arbitrary point π0 ∈ D, say with π0 (x) = ax + r. Also let t1 a fixed root of the
equation t2 + t + 1 = 0 over GF (q), where two distinct roots to this quadratic are guaranteed
to exist in GF (q) by Corollary 19a (see the appendix).
For the claim which follows, let the vertices of D be listed as π0 π1 π2 · · · πk π0 , k ≥ 3, as we
go cyclically around D; so πi πi+1 , 0 ≤ i ≤ k − 1, and πk π0 are the edges of D.
Consider the subgroup Q = {x + b : b ∈ GF (q)} of AGL(1, q). Clearly |Q| = q, and for each
h ∈ GF (q), h 6= 0, Q has the coset hxQ = {hx + b : b ∈ GF (q)}, which we abbreviate by Qh .
Claim A: a) πi ∈ Qati1 for each 0 ≤ i ≤ k, and also π0 ∈ Qatk+1 .
1
π0 (F )−F

b) If πi πi+1 ∈ E(CA (q)) for 0 ≤ i ≤ k − 1, then tj1 = (−1)j πj (F )−F
for 1 ≤ j ≤ k.

6
c) If the graph induced by π0 , π1 , π2 , · · · , πk in CA (q) is the cycle Ck+1 ; that is, πi πi+1 ∈ E(CA (q))
for 0 ≤ i ≤ k − 1 and π0 πk ∈ E(CA (q)), then tk+1 1 = (−1)k+1 .
Proof of Claim A: Consider a) of the claim. By the definition of edges in H(q) we may take π1 to
be of the form π1 (x) = at1 x+c (otherwise change t1 to t11 ), so π1 ∈ Qat1 . Proceeding inductively,
assume that πi ∈ Qati1 for i ≤ j < k. By part a) of the theorem πj+1 has exactly two neighbors
in H(q), one of them in Qatj−1 and the other in Qatj+1 . Now πj−1 ∈ Qatj−1 by induction, and
1 1 1
πj+1 6= πj−1 since k ≥ 3. It follows that πj+1 ∈ Qatj+1 , completing the inductive step. As in the
1
inductive step we just argued, π0 must be the neighbor of πk in Qatk+1 , completing part a).
1
Now consider b) of the claim. Writing π1 (x) = bx + c (so b = at1 by part a) of the claim) and
using π0 π1 ∈ E(CA (q)), we apply the first and third equations of (1),
with ππ00(Fand π1 playing
b i−F )−F

the roles of σ and π respectively. We then get t1 = a = − j−F = − π1 (F )−F . For the
inductive step suppose the claimed equation for tj1 holds for 0 ≤ j ≤ r < k. Then let πr and
πr+1 play the roles of σ and π again in the first and third equations of

(1). Now applying the
r+1 r πr (F )−F
r π0 (F )−F r+1 π0 (F )−F


inductive hypothesis we get t1 = t1 t1 = − πr+1 (F )−F (−1) πr (F )−F = (−1) πr+1 (F )−F
,
as required.
For c) of the claim, we apply b) with k + 1 playing the role of k,
by regarding π0 as πk+1 with
k+1 k+1 π0 (F )−F k+1
edge πk πk+1 ∈ E(CA (q)). It follows that t1 = (−1) π0 (F )−F
= (−1) . QED
We now complete the proof of part b) of the theorem, using in what follows the facts that
t3 −1
t31= 1 and t21 6= 1. The first of these follows from 0 = 1 + t1 + t21 = t11 −1 , and from t1 6= 1 by
Lemma 3c2. The second follows from t31 = 1 and t1 6= 1.
Suppose to the contrary that our component D of H(q) is isomorphic to C3s+p , where p = 1
or 2 for some integer s ≥ 1. Take an arbitrary point π0 in D, with π0 (x) = ax + r. Assume first
that p = 1, and let π0 , π1 , · · · , π3s be the vertices of D listed cyclically as we traverse D, ie.,
with πi πi+1 ∈ E(H(q)), π3s π0 ∈ E(H(q)), and wlog π1 (x) = at1 x + r. Then by Claim Aa) we
have π0 ∈ Qat3s+1
1
= Qat1 , contradicting π0 ∈ Qa . Now suppose p = 2. Then the vertices of D
can be written in cyclic order π0 , π1 , · · · , π3s+1 . Again by Claim Aa) we get π0 ∈ Qat3s+2
1
= Qat21 ,
2
again contradicting π0 ∈ Qa since t1 6= 1. This completes the proof of part b).
Consider part c). Using part b), we may suppose to the contrary that D ∼ = C3s , where s
is odd. Then we have D ∼ = C6r+3 for suitable r. By Claim Ac and recalling t31 = 1 we obtain
1 = t6r+3
1 = (−1)6r+3 = −1, a contradiction. Thus s is even, so D ∼ = C6r , as required for c).

Corollary 5 Let q be a prime power with q ≡ 1 (mod 3) and q ≥ 7. Then

a) if q is odd, then M (q − 1, q − 3) ≥ (q 2 − 1)/2, and
b) if q is even, then M (q − 1, q − 3) ≥ (q − 1)(q + 2)/3.

Proof. From the discussion preceding Lemma 3, we have the lower bound M (q − 1, q − 3) ≥ |I|,
where I is any independent set in CA (q). We now find such an independent set. For any subgraph
H of H(q), let αC (H) be the maximum size of a set of vertices in H which is independent in
CA (q).
Let Y be the set of isolated points of CA (q). Let D any connected component of H(q) − Y .
Since by Theorem 4a every connected component of H(q) is a cycle, it follows that D must be
either a cycle or a path.
For part a), assume first that q is odd. We will show that αC (D) ≥ d |D|
2
e. Suppose first that
D is a cycle. If D is also a connected component of CA (q), then D ∼ = C6r by Theorem 4c. Then
immediately αC (D) = 3r = |D| 2
, as required. Still assuming D is a cycle, suppose D is not a

7
component of CA (q). So D ∼ = C3s by Theorem 4b, and D contains an edge e with e ∈ / CA (q).
Then D − e is the path P3s on 3s vertices. This P3s has an independent set in H(q) of size d 3s 2
e,
|D|
and that set is also independent in CA (q) since CA (q) ⊆ H(q). Thus αC (D) ≥ d 2 e.
Still with q odd, suppose D is a path. Then again D contains an independent set in H(q) of
size d |D|
2
e, and hence also αC (D) ≥ d |D|
2
e. P
Thus altogether we obtain αC (H(q)) ≥ |Y | + αC (D), where the sum is over all connected
components D of H(q)−Y . So using Lemma 3b we get αC (H(q)) ≥ q −1+ 12 (q(q −1)−(q −1)) =
1 2
2
(q − 1), as required.
Next consider b), so q is even. Now αC (D) is minimized, given that D ∼ = C3s , when D ∼ = C3
|D|
and D ⊆ CA (q). In that case we get αC (D) = 1 = 3
. By a similar estimate as above we then
1


get αC (H(q)) ≥ q − 1 + 3 q(q − 1) − (q − 1) = (q − 1)(q + 2)/3.

The lower bounds in this corollary should be compared to the lower bound M (q, q − 2) ≥ q 2
for prime powers q 6≡ 2(mod 3), derived by using permutation polynomials [8].

3 The contraction graph for P GL(2, q)

Let q be power of a prime. The permutation group P GL(2, q) is defined as the set of one to one
functions σ : GF (q) ∪ {∞} → GF (q) ∪ {∞}, under the binary operation of composition, given
by
ax + b
{ σ(x) = | a, b, c, d ∈ GF (q), ad 6= bc, x ∈ GF (q) ∪ {∞} }. (5)
cx + d
Here σ(x) is computed by the rules:
ax+b
1. if x ∈ GF (q) and x 6= −(d/c), then σ(x) = cx+d
,

2. if x ∈ GF (q) and x = −(d/c), then σ(x) = ∞,

3. if x = ∞, and c 6= 0, then σ(x) = a/c, and

4. if x = ∞, and c = 0, then σ(x) = ∞.

We regard P GL(2, q) as a permutation group acting on the set GF (q) ∪ {∞} of size q + 1 via
the one to one map x 7→ σ(x). One can show that |P GL(2, q)| = (q + 1)q(q − 1), and it is well
known that P GL(2, q) is sharply 3-transitive in its action on GF (q) ∪ {∞} (see [23] for a proof).
By Theorem 1, it follows that hd(P GL(2, q)) = q − 1, and M (q + 1, q − 1) = |P GL(2, q)| =
(q + 1)q(q − 1).
Take a fixed ordering of GF (q)∪{∞} with ∞ as final symbol, say x1 , x2 , · · · , xq , ∞ where the
xi are the distinct elements of GF (q). Then any element π ∈ P GL(2, q) is identified with the
length q +1 string π(x1 )π(x2 )π(x3 ) · · · π(xq )π(∞), which again we call the image string of σ. For
any such π ∈ P GL(2, q) the permutation π 4 on GF (q) ∪ {∞} is defined as in the introduction,
where F = ∞ is the distinguished element of GF (q) ∪ {∞} in that definition. As an example,
4
if π = a∞bcde, then π 4 = aebcd∞, and π− = aebcd. Further, for any subset R ⊂ P GL(2, q),
4 4
the sets R , and R− are defined in the same way as in the introduction, with F = ∞. Since
hd(P GL(2, q)) = q − 1 = q + 1 − 2, the image strings of any two elements of P GL(2, q) agree
in at most two positions. It follows from Lemma 2a that for any π, σ ∈ P GL(2, q) we have
hd(π 4 , σ 4 ) ≥ hd(π, σ) − 3 ≥ q − 4. That is, π 4 and σ 4 can agree in at most 5 positions; up to 2
occurring from the original π and σ, and up to 3 more occurring from the π 4 and σ 4 operation.
We show that if q 6≡ 1(mod 3), then hd(P GL(2, q)4 ) ≥ q − 3. Set G = P GL(2, q) for brevity.
Assuming otherwise and using hd(G) = q −1, we have by the immediately preceding inequalities

8
(based on Lemma 2a) that hd(G4 ) = q − 4 and for some π, σ ∈ H we have hd(π, σ) = q − 1
and hd(π 4 , σ 4 ) = hd(π, σ) − 3. Let x and y be the two coordinates of agreement between
π and σ, and let Gx,y be the stabilizer subgroup for {x, y}; that is, the subgroup of G fixing
both x and y. It is straightforward to check that G∞ ∼ = AGL(1, q), and by transitivity of G also
Gx ∼ = AGL(1, q) for any x ∈ GF (q)∪{∞}. Thus Gx,y is the 1-point stabilizer of AGL(1, q), with
AGL(1, q) acting transitively on GF (q) as in the previous section. By standard permutation
group theory, if a group H acts transitively on a set Ω, and Hz is the stabilizer of z in H, for
|H|
z ∈ Ω, then |H z|
= |Ω|. Thus q = q(q−1)
|Gx,y |
. Hence |Gx,y | = q − 1. By Lemma 2b, πσ −1 contains a
3-cycle in its cycle factorization. So 3 divides |Gx,y | = q − 1, a contradiction.
So with q 6≡ 1(mod 3) we have that G4 − is a permutation array on q symbols satisfying
hd(G− ) = hd(G ) ≥ q − 3 using Lemma 2c. Also by Lemma 2c we have |G4
4 4 4
− | = |G | = |G| =
4
(q + 1)q(q − 1) since q ≥ 13. So trivially we get the bound M (q, q − 3) ≥ |G− | = (q + 1)q(q − 1).
When q = 2k , still with q 6≡ 1(mod 3), the improvement M (q, q − 3) ≥ (q + 2)q(q − 1) as well as
the bound M (q, q − 4) ≥ 31 q(q − 1)(q 2 + 3q + 8) are given in [8] using permutation polynomials.
Thus in this section, we restrict ourselves to the case q ≡ 1(mod 3), q an odd prime power,
and for technical reasons we take q ≥ 13.
The plan will be similar in some respects to the one we used in the previous section. That is,
for a certain set I ⊂ P GL(2, q) we will find a permutation array I−4 ⊂ P GL(2, q)4 − on q symbols
4 4
with hd(I− ) ≥ q − 3, thus obtaining the lower bound on M (q, q − 3) ≥ |I− |. This set I will be
an independent set in a graph which we now define.
Define the contraction graph CP (q) by V (CP (q)) = P GL(2, q), and E(CP (q)) = {πσ :
hd(π 4 , σ 4 ) = q − 4}. So edges of CP (q) correspond to pairs π, σ for which hd(π 4 , σ 4 ) achieves
its least possible value of q − 4, occurring when π 4 and σ 4 agree in 5 postions. Thus any inde-
pendent set I in CP (q) satisfies hd(I 4 ) ≥ q − 3. By Lemma 2c, we get hd(I−4 ) = hd(I 4 ) ≥ q − 3,
while |I−4 | = |I 4 | = |I|, with the last equality following from hd(I 4 ) = q − 3 > 3 since q ≥ 13.
We are thus reduced to finding an independent set I in CP (q), from which M (q, q − 3) ≥ |I|
follows.
We now study the structure of CP (q). Consider π ∈ P GL(2, q), say with π(x) = ax+b cx+d
, where
x ∈ GF (q) ∪ {∞}. Now c = 0 is equivalent to π(∞) = ∞. Thus π is an isolated point in CP (q).
Later we show that π is isolated in CP (q) if and only if π(∞) = ∞. Suppose then that c 6= 0.
ax+b
a
x+ cb
a
(x+ dc )+ bc−ad bc−ad
Then some manipulation shows that π(x) = cx+d
= c
x+ dc
= c
x+ dc
c2
= ac + c2
x+ dc
. So to study
the structure of CP (q) apart from its isolated points, we are reduced to considering those vertices
r
of CP (q) which can be expressed as π(x) = K + x−i , for suitable elements K, r, i ∈ GF (q), r 6= 0,
which depend on π.
r s
Lemma 6 Let π, σ ∈ P GL(2, q) with π(x) = a + x−i , σ(x) = b + x−j
iwith r, s 6= 0. Then
4 4
hd(π , σ ) = hd(π, σ) − 3 ⇐⇒ (b − a)(j − i) = r and r = s .

Proof. =⇒ : By assumption we have π(∞) = a and π(i) = ∞, together with σ(j) = ∞ and
σ(∞) = b. By lemma ??a the only coordinates of either π or σ whose values are affected by
the 4 operation are the 3 coordinates π −1 (∞) = i, σ −1 (∞) = j, and ∞. So the assumption
hd(π 4 , σ 4 ) = hd(π, σ) − 3 implies that σ(i) = a and π(j) = b. Thus we get π(j) = a + j−i
r
= b,
yielding (b − a)(j − i) = r as required. Now interchanging the roles of π and σ in this argument,
s
specifically, using σ(i) = b + i−j = a, we get (a − b)(i − j) = s, so also r = s.
⇐= : Again by assumption we have π(∞) = a, σ(∞) = b, π(i) = ∞, σ(j) = ∞, and
(b − a)(j − i) = r . To prove hd(π 4 , σ 4 ) = hd(π, σ) − 3, it remains only to show that π(j) = b
and σ(i) = a. For simplicity we let r = s = 1, since the argument does not depend on

9
 ... ... ... Bgq−1 = B1
(1, aq ) (1, a3 ) (1, a2 ) (1, a1 )
..
.
... ... ... Bg i
(g i , aq ) (g i , a3 ) (g i , a2 ) (g i , a1 )
..
.
... ... ... Bg
(g, aq ) (g, a3 ) (g, a2 ) (g, a1 )
..
.
... ... ... B0
(0, aq ) (0, a3 ) (0, a2 ) (0, a1 )

Figure 1: The graph P1 , partitioned into levels B0 and Bgi , 1 ≤ i ≤ q − 1.

1
r = s. Solving for b in (b − a)(j − i) = 1 we get b = j−i
+ a = π(j). Solving for a we get
1 1
a = b − j−i = b + i−j = σ(i), as required.

Lemma 7 Let q = pm , where p is an odd prime, with q ≡ 1(mod 3), q ≥ 13. Let π, σ ∈
r
P GL(2, q), say with π(x) = a + x−i s
, σ(x) = b + x−j , with r, s 6= 0. Then hd(π 4 , σ 4 ) =
hd(π, σ) − 3 ⇐⇒ πσ ∈ E(CP (q)).

Proof. ⇐=: By definition of edges in CP (q) and Lemma 2a we have q − 4 = hd(π 4 , σ 4 ) ≥

hd(π, σ)−3. Now since q −1 ≤ hd(π, σ) ≤ q +1, equality is forced together with hd(π, σ) = q −1.
This yields hd(π 4 , σ 4 ) = hd(π, σ) − 3.
=⇒ : By the assumption hd(π 4 , σ 4 ) = hd(π, σ) − 3 and hd(π, σ) ≥ q − 1 we are reduced to
showing that hd(π, σ) = q − 1; that is, that π and σ already agree in two coordinates.
r r
By assumption and Lemma 6 we have r = s, so write π(x) = a + x−i and σ(x) = b + x−j , for
a, b, i, j, k ∈ GF (q) with r 6= 0. Note also i 6= j by Lemma 6 since r 6= 0.
We now derive a quadratic equation over GF (q) whose distinct roots are the coordinates of
agreement between π and σ. Since hd(π 4 , σ 4 ) = hd(π, σ)−3, by Lemma 6 we have (b−a)(j−i) =
r
r. Thus b = j−i + a. Now we set π(x) = σ(x) to find the possible coordinates x at which π and
σ agree, understanding that x can be neither i nor j since π and σ can have no agreements in
r 1 1 1
coordinates i, j and ∞. Substituting j−i + a for b and simplifying we obtain x−i − x−j = j−i .
i−j 1 2 2
Hence (x−i)(x−j) = j−i , and we get the quadratic x − (i + j)x + ij + (i − j) = 0. By Corollary
19b there are two distant roots√to this equation, √ giving the two coordinates
√ of agreement
√ for π
and σ as follows; x1 = 12 [i(1 + −3) + j(1 − −3)], and x2 = 21 [i(1 − −3) + j(1 + −3)].
Hence by our reduction at the beginning of the proof it follows that πσ ∈ E(CP (q), as
required.

The preceding two Lemmas yield the following.

Corollary 8 Let q = pm , where p is an odd prime, with q ≡ 1(mod 3), q ≥ 13.

r s
a) Let π, σ ∈ P GL(2, q), say with π(x) = a + x−i , σ(x) = b + x−j , r, s 6= 0. Then πσ ∈
E(CP (q)) ⇐⇒ r = s and (b − a)(j − i) = r.
b) π ∈ P GL(2, q) is an isolated point in CP (q) ⇐⇒ π(∞) = ∞.

10
Proof. Part a) follows immediately from Lemmas 6 and 7.
For part b), suppose first that π(∞) = ∞. Then by definition of π 4 , with ∞ playing the role
of F , we have π 4 = π. Suppose to the contrary that π has a neighbor σ in CP (q). By Lemma
7 we have hd(σ 4 , π 4 ) = hd(σ, π) − 3. If σ 4 = σ, then hd(σ 4 , π 4 ) = hd(π, σ), contradicting
Lemma 7. So assume σ 4 6= σ, and thus σ(∞) 6= ∞. Then the only coordinates at which σ 4 and
σ could possibly disagree are σ −1 (∞) and ∞. Thus hd(σ 4 , π 4 ) ≥ hd(π, σ) − 2, contradicting
Lemma 7.
Conversely, suppose to the contrary that π is isolated and π(∞) = x 6= ∞. Let i = π −1 (∞),
and let j be any coordinate with j ∈ / {i, ∞}, and let π(j) = y. Then by sharp 3-transitivity of
P GL(2, q) we can find an element σ ∈ P GL(2, q) satisfying σ(j) = ∞, σ(i) = x, and σ(∞) = y.
Then we get hd(σ 4 , π 4 ) = hd(σ, π) − 3. So by Lemma 7 we have πσ ∈ E(CP (q)), contradicting
π being isolated.

The next two theorems, which use the preceding Lemma, tell us more about CP (q). For
any S ⊂ CG(q), we let [S] be the subgraph of CP (q) induced by S; that is, V ([S]) = S and
E([S]) = {πσ : π, σ ∈ S, πσ ∈ E(CP (q))}. When r is fixed by context, we denote a vertex
r
v = a + x−i by the abbreviation (i, a).
Consider the partition of P GL(2, q) given by P GL(2, q) = ∪r6=0 Pr , where for r ∈ GF (q) with
r
r 6= 0, Pr = {a + x−i : a, i ∈ GF (q)}, so |Pr | = q 2 . Further consider the partition of Pr given by
r
Pr = ∪i∈GF (q) Bi (r), where Bi (r) = {a + x−i : a ∈ GF (q)}.

Theorem 9 Let q = pm , where p is an odd prime, with q ≡ 1(mod 3), q ≥ 13. Then the
following hold in the graph CP (q).
a) For any r 6= s, r, s 6= 0, we have [Pr ] ∼
= [Ps ].
b) For any r 6= 0 and i 6= j, [Bi (r) ∪ Bj (r)] is a perfect matching, which matches Bi (r) to Bj (r].
c) For any r 6= 0, the subgraph [Pr ] is regular of degree q − 1.
d) For any nonisolated vertex v ∈ CP (q), [N (v)] is a disjoint union of cycles, where N (v) is the
set of neighbors of v in CP (q).
1
Proof. For a), consider for any r ∈ GF (q), r 6= 0, the map ϕ : P1 → Pr given by ϕ(a + x−i )=
r 1 1
a + x−ri . Let v, w ∈ P1 , say with v = a + x−i and w = b + x−j . Then vw ∈ E([P1 ]) ⇔
(b − a)(j − i) = 1 ⇔ (b − a)(rj − ri) = r ⇔ ϕ(v)ϕ(w) ∈ E([Pr ]) . Thus ϕ is a graph
isomorphism, and since r was arbitrary, it follows that for any s 6= 0 we have [Pr ] ∼ = [P1 ] ∼
= [Ps ].
Consider b). Fix r, and consider any two points (i, a) and (j, b) of Pr . By Corollary 8
we have (i, a)(j, b) ∈ E(CP (q)) if and only if i 6= j and (b − a)(j − i) = r in GF (q). Let
Hij = [Bi (r) ∪ Bj (r)] for i 6= j. Note there can be no edge in Hij of the form (i, a)(i, b) since
(b − a)(i − i) = 0 6= r, and similarly no edge of the form (j, a)(j, b). Now given (i, a) ∈ Bi (r), a
point (j, b) ∈ Bj (r) is a neighbor of (i, a) if and only if (b − a)(j − i) = r by Corollary 8.
Thus for this fixed i and j we can uniquely determine b by the equation b = r(j − i)−1 + a,
showing that (j, b) is the only neighbor of (i, a) in Bj (r). A symmetric argument shows that
each point in Bj (r) has a unique neighbor in Bi (r). Thus E(Hij ) is a perfect matching, which
matches Bi (r) to Bj (r).
For c), let v ∈ CP (q), say with v ∈ Bi (r) ⊂ Pr for some r 6= 0. By Corollary 8, any neighbor
of v in CP (q) must also lie in Pr . By part b), the neighbors of v are in one to one correspondence
with the sets Bj (r), j 6= i, j ∈ GF (q). Thus v has exactly |GF (q)| − 1 = q − 1 neighbors in
CP (q).
For d), take v ∈ CP (q), and by the isomorphism of subgraphs [Pr ] from part b), we can take
v = (i, a) ∈ P1 . By Corollary 8 we have N (v) ⊂ P1 . It suffices to show that [N (v)] is regular of
degree 2. Let (j, b) ∈ N (v), so j 6= i by part b). Now any neighbor (k, c) of (j, b) in N (v) must

11
 Bgq−1

..
.

Bg i

..
.

Bg

..
.

B0

Figure 2: Perfect matching between any two levels of P1 .

lie in N ((i, a)) ∩ N ((j, b)). So to show that (j, b) has degree 2 in [N (v)], it suffices to show that
(k, c) ∈ P1 satisfies (k, c) ∈ N ((i, a)) ∩ N ((j, b)) if and only if k is a root in GF (q) of a quadratic
equation over GF (q) having two distinct roots in GF (q).
Suppose first that (k, c) ∈ N ((i, a)) ∩ N ((j, b)). By Corollary 8 we must have the equations
(c − a)(k − i) = 1, (b − c)(j − k) = 1, (b − a)(j − i) = 1.
Using the second and third equations we get c = (j − i)−1 − (j − k)−1 + a, and from the first
equation c = (k − i)−1 + a. Setting these two expressions for c equal we obtain (k − i)−1 + (j −
k)−1 = (j − i)−1 . Some simplification leads to the quadratic k 2 − k(i + j) + ij + (j − i)2 = 0
with coefficients over GF (q) and unknown k. By Corollary 19b from √ the Appendix,
√ we see
that that there are two distinct solutions for k; namely k1 = 21 [i(1 + −3) + j(1 − −3)], and
√ √
k2 = 12 [i(1 − −3) + j(1 + −3)].
Conversely suppose that k is one of the two distinct solutions of k 2 −k(i+j)+ij +(j

1−i) =
2
0.
1 1 1
2 2


Then (k − i)(j − k) = −k + k(i + j) − ij = (j − i) , and using (k−i)(j−k) = j−i k−i + j−k ,
1 1 1 1
one can derive k−i + j−k = j−i . Now set c = k−i + a, so immediately we get (c − a)(k − i) = 1.
1
Since (i, a) and (j, b) are neighbors we have (b − a)(j − i) = 1, so b = j−i + a. It follows that
1 1 1 1
c = k−i + a = j−i − j−k + a = b − j−k . Hence we get (b − c)(j − k) = 1. Thus the three
equations (c − a)(k − i) = 1, (b − c)(j − k) = 1, and (b − a)(j − i) = 1 hold, showing that
(k, c) ∈ N ((i, a)) ∩ N ((j, b)) by Corollary 8.
Note that once k is determined (as one of the two distinct roots), then the point (k, c) is
uniquely determined by the perfect matching between Bk (1) and Bi (1) (or Bj (1)). Thus we
obtain that an arbitrary point (j, b) ∈ N (v) has exactly two neighbors in N (v), completing d).

To round out the structure of CP (q) we consider the connected components of CP (q).
Theorem 10 Let q = pm , where p is an odd prime, with q ≡ 1(mod 3), q ≥ 13. Then the
connected components of CP (q) are as follows.
1) the isolated points - these are of the form π(x) = ax + b, a 6= 0, and there are q(q − 1) of
them,
2) the graphs [Pr ] induced by the sets Pr .

12
Proof. By Corollary 8b we have that π ∈ P GL(2, q) is an isolated point in CP (q) if and only if
π(∞) = ∞. This is equivalent to π(x) = ax+b, a 6= 0 and there q(q −1) such points, completing
part 1).
r
The remaining permutations are all of the form π(x) = a + x−i for suitable a, r, i ∈ GF (q)
with r 6= 0 as shown earlier. Hence it suffices to analyze the connected component structure of
[∪r6=0 Pr ]. By Corollary 8 and Theorem 9a, to prove part 2) it suffices to prove that any one of
the [Pr ], say [P1 ], is connected.
Recall the partition P1 = ∪i∈GF (q) Bi (1) defined above, and from now on we abbreviate Bi (1)
by Bi . Let g by a generator of the multiplicative cyclic subgroup of nonzero elements in GF (q).
Then we can write this partition as P1 = B0 ∪(∪1≤k≤q−1 Bgk ). We regard the sets in this partition
as “levels” of CP (q); where B0 is level 0 and B(g k ) is level k, 1 ≤ k ≤ q − 1. See Figure 1 for an
illustration of P1 from this viewpoint. By Theorem 9b the subgraph of [P1 ] induced by any two
levels has edge set which is a perfect matching, as illustrated in Figure 2.
First we observe that to show that [P1 ] is connected it suffices to show that any two vertices
in B0 are joined by a path in [P1 ]. For if that was true, then we can find a path in [P1 ] from
(0, 0) to any vertex w ∈ P1 (thus showing connectedness of [P1 ]) as follows. If w ∈ B0 we are
done by assumption. So suppose w ∈ / B0 , say with w ∈ B(g k ). Let s be the unique neighbor in
B0 of w under the perfect matching E([B0 ∪ B(g k )]). Let P be the path from (0, 0) to s in [P1 ]
which exists by assumption. Then P followed by the edge sw is a walk joining (0, 0) to w, so P
contains a path from (0, 0) to w.
By Theorem 9b there is a (unique) path in [P1 ] starting at (0, 0) and passing through levels
1, 2, · · · , q − 1 in succession. Let (0, 0) − (g, α1 ) − (g 2 , α2 ) − ... − (g q−1 , αq−1 ) be this path, for
suitable αk ∈ GF (q). For k ≥ 1 let (0, βk ) ∈ B0 be the unique neighbor in level 0 of the vertex
(g k , αk ) in level k. This path and the points (0, βk ) are illustrated Figure 3. Our first step is to
obtain the values of αk and βk .
..
4
.
(g , α4 )
Bg 4

(g 3 , α3 )
Bg 3

(g 2 , α2 )
Bg 2

(g 1 , α1 )
Bg

B0
(0, β4 ) (0, β3 ) (0, β2 ) (0, 0) = (0, β1 )

Figure 3: The path (0, 0) − (g, α1 ) − (g 2 , α2 ) − · · · − (g q−1 , αq−1 ), and (0, βi ) is the neighbor of
(g i , αi ) at level 0.

Claim 1: We have

13
 g k−1 +g k−3 +g k−4 +···+g+1
a) α1 = g1 , α2 = 1
g−1
, and αk = (g−1)g k−1
for k ≥ 3.
(g 2 −g+1)(1+g+g 2 +g 3 +···+g k−2 )
b) β1 = 0, and βk = g k (g−1)
for k ≥ 2.
Proof of Claim 1: We repeatedly use the fact, proved earlier, that if (r, a) and (s, b) are adjacent
vertices in the contraction graph CP (q), then (s − r)(b − a) = 1.
For part a), since (0, 0) − (g, α1 ) is an edge in CP (q) we have (α1 − 0)(g − 0) = 1, so
α1 = g1 . Since (g, α1 ) − (g 2 , α2 ) is an edge we have (α2 − g1 )(g 2 − g) = 1, yielding α2 = g−1
1
,
g +1 2
1
and similarly (α3 − g−1 )(g 3 − g 2 ) = 1, yielding α3 = (g−1)g 2 . Now for k ≥ 3 we proceed by

induction, having proved the base case k = 3. Since (g , αk ) − (g k−1 , αk−1 ) is an edge, we have
k

(αk − αk−1 )(g k − g k−1 ) = 1. Solving for αk and applying the inductive hypothesis to αk−1 , we
k−2 k−4 +g k−5 +···+g+1
obtain αk = gk −g1 k−1 + g +g (g−1)g k−2 , which after simplification yields the claim.
For part b), we have β1 = 0 since (0, 0) − (g, α1 ) is an edge. Since (g 2 , α2 ) − (0, β2 ) is an
1
edge, we have ( g−1 − β2 )(g 2 − 0) = 1, and solving for β2 and simplifying we get the claim for
k = 2. Consider now k ≥ 2. The existence of edge (g k , αk ) − (0, βk ) gives (αk − βk )g k = 1, so
k−1 k−3 +g k−4 +···+g+1
βk = αk − g1k . Using the formula for αk from part a), we have βk = g +g (g−1)g k−1 − g1k =
g k +g k−2 +g k−3 +···+g 2 +1 (g 2 −g+1)(1+g+g 2 +g 3 +···+g k−2 )
(g−1)g k
= g k (g−1)
. QED
Claim 2: We have |{βk : 1 ≤ k ≤ q − 1}| = q − 1; that is, the βk , 1 ≤ k ≤ q − 1, are pairwise
distinct.
Proof of Claim 2: In applying Claim 1, we note first that g could have been chosen so as not to
be a root of x2 − x + 1 = 0 as follows. The number of roots in GF (q) to this quadratic is at
most 2. Now the number of generators in the multiplicative cyclic group GF (q) − {0} of order
q − 1 is the euler totient function φ(q − 1), defined as the number of integers 1 ≤ s ≤ q − 1
which are relatively prime to q − 1. Since q is an odd prime power with q ≥ 13, we know that
φ(q − 1) > 2, so such a g exists.
We show that for for any pair j, k with 1 ≤ j < k ≤ q − 1 we have βk 6= βj .
Consider first the case j = 1. Since β1 = 0, we need to show that βk 6= 0 for 2 ≤ k ≤ q − 1.
2 2 +g 3 +···+g k−2 )
Supposing the contrary and applying Claim 1b we get (g −g+1)(1+g+g
k
g (g−1)
= 0. Canceling
2 k−1
the nonzero factor ggk −g+1
(g−1)
on the left side, we get 0 = (1 + g + g 2 + g 3 + · · · + g k−2 ) = g g−1−1 . This
implies that g k−1 − 1 = 0, which is impossible since k − 1 ≤ q − 2 while g, being a generator of
the group, must have order q − 1.
So now suppose that j ≥ 2. Assuming the contrary that βk = βj and applying Claim 1b, we
get after simplification that 1 + g + g 2 + g 3 + · · · + g k−2 = g k−j (1 + g + g 2 + g 3 + · · · + g j−2 ) =
k−j
g k−j + g k−j+1 + · · · + g k−2 . Thus we have 0 = 1 + g + g 2 + · · · + g k−j−1 = g g−1−1 . So g k−j = 1,
which is impossible since k − j ≤ q − 3, while g has order q − 1. QED
We introduce some notation in preparation for the rest of the argument. Let Z = {(0, βk ) :
1 ≤ k ≤ q − 1} ⊂ B0 . Since |B0 | = q, by Claim 2 we have |B0 − Z| = 1, and we let
u be the unique vertex of B0 − Z. Further for any subset T of vertices in CP (q), we let
N (T ) = {v ∈ CP (q) : v ∈ / T, vt ∈ E(CP (q)) for some t ∈ T } be the neighbor set of T in CP (q).
Recall also that [T ] denotes the subgraph of CP (q) induced by T .
Claim 3: [Z ∪ N (Z)] is a connected subgraph of CP (q).
Proof of Claim 3: For brevity let H = [Z ∪ N (Z)]. Since β1 = 0 we have (0, 0) ∈ Z. Thus it
suffices to show that for any w ∈ V (H) there is a path in H joining (0, 0) to w.
Suppose first that w ∈ Z, so w = (0, βk ) for some k. Observe that (g i , αi ) ∈ N (Z) for
all i by definition. So the path (0, 0) − (g, α1 ) − (g 2 , α2 ) − ... − (g k , αk ) followed by the edge
(g k , αk ) − (0, βk ) is path in H joining (0, 0) to w.

14
 Next suppose w ∈ N (Z), say with w adjacent to (0, βk ) ∈ Z. Then the path (0, 0) − (g, α1 ) −
(g 2 , α2 ) − ... − (g k , αk ) followed by the length 2 path (g k , αk ) − (0, βk ) − w is a walk in H joining
(0, 0) to w, and this walk contains the required path. QED
We can now complete the proof of the theorem. Let H 0 = [{u} ∪ N (u)]. First we observe
that V (H) ∩ V (H 0 ) = ∅ as follows. Suppose not, and let z ∈ V (H) ∩ V (H 0 ), say with z ∈ B(g k ),
noting that k ≥ 1 since each level, in particular B0 , is an independent set in [P1 ]. Then z has
two distinct neighbors in B0 ; namely u and (0, βj ), for some 1 ≤ j ≤ q − 1. This contradicts the
fact that the edge set of [B(g k ) ∪ B0 ] is a perfect matching between the levels B(g k ) and B0 by
Theorem 9b. Thus V (H) ∩ V (H 0 ) = ∅.
Observe that H 0 is connected (because H 0 contains the spanning star K1,q−1 with u as center)
and H is connected by Claim 3, while V (H) ∪ V (H 0 ) = P1 and V (H) ∩ V (H 0 ) = ∅. Thus to
prove that [P1 ] is connected we are reduced to showing that there is an edge vw ∈ E([P1 ]) with
v ∈ H 0 and w ∈ H. Suppose no such edge exists. Since [P1 ] is (q − 1)-regular by Theorem 9c,
it follows that H 0 is a q − 1 regular graph on q vertices. Thus H 0 = Kq . Hence [N (u)] = Kq−1 .
But this is a contradiction since by Theorem 9d the neighborhood of any nonisolated point in
CP (q) is regular of degree 2, while [N (u)] is regular of degree q − 2 > 2 since q ≥ 13.

We can now obtain our independent set in CP (q) as a consequence of our previous results
and the following theorem of Alon [1].

Theorem 11 [1] Let G = (V,E) be a graph on N vertices with average degree t ≥ 1 in which
for every vertex v ∈ V the induced subgraph on the set of all neighbors of v is r-colorable. Then
c N
the maximum size α(G) of an independent set in G satisfies α(G) ≥ log(r+1) t
log t, for some
absolute constant c.

Corollary 12 Let q be a power of an odd prime p, with q ≡ 1(mod 3), q ≥ 13.

a) α(CP (q)) ≥ Kq 2 log q for some constant K.
b) M (q, q − 3) ≥ Kq 2 log q for some constant K.

Proof. Consider a). By Corollary 8a there is no edge between any two subgraphs [Pr ] and [Ps ]
for r 6= s. Since there are q such subgraphs, and by Theorem 9a) they are pairwise isomorphic,
it suffices to show that α(P1 ) ≥ Kq log q for some constant K.
We now apply Alon’s theorem to the subgraph [P1 ] of CP (q). Now [P1 ] is (q − 1)-regular by
Theorem 9c, and has q 2 points. Since the neighborhood of every point is a disjoint union of
cycles by Theorem 9d, this neighborhood must be 3-colorable. It follows by Alon’s theorem that
q2
[P1 ] contains an independent set of size logc 4 q−1 log(q − 1) ∼ Kq log q, for some constant K.
For b), let I be an independent set in CP (q) of size Kq 2 log q for suitable constant K, guar-
anteed to exist by by part a). Then by the reduction made in the discussion preceding Lemma
6 we have M (q, q − 3) ≥ |I| ≥ Kq 2 log q.

4 Lower bounds for M (n, d) for special n, d via the Math-

ieu groups
In this section we consider the Mathieu groups M11 , M12 , M22 , M23 , M24 , discovered by E. Math-
ieu in 1861 and 1873. These permutation groups are the earliest known example of sporadic
simple groups. See [10], [6], or [24] for a discussion of their construction. These groups act on
11, 12, 22, 23, 24 letters respectively, with M11 being a 1 point stabilizer of M12 , while M23 and
M22 are 1 and 2 point stabilizers of M24 respectively.

15
 In this section we apply the contraction operation to these permutation groups to obtain new
permutation arrays, with resulting lower bounds for M (n, d) for suitable n and d.
Since M12 is sharply 5-transitive we have by Theorem 1 that hd(M12 ) = 8 and M (12, 8) =
|M12 | = 95040. Similarly since M11 is sharply 4-transitive we have M (11, 8) = |M11 | = 7920.
For M24 we do not have sharp transitivity. But observe that for any permutation group G acting
on some set, and three elements π, σ, τ ∈ G, we have hd(π, σ) = hd(πτ, στ ) = hd(τ π, τ σ). Thus
hd(G) =min{hd(1, σ) : σ ∈ G}. From the set of disjoint cycle structures of elements of M24
(available at [29]) we find that the largest number of 1-cycles in the disjoint cycle structure of
any nonidentity element of M24 is 8. Thus hd(M24 ) = 24 − 8 = 16, and from the stabilizer
relation also hd(M23 ) = hd(M22 ) = 16. We thus obtain M (24, 16) ≥ |M24 | = 24, 423, 040,
M (23, 16) ≥ |M23 | = 10, 200, 960, and M (22, 16) ≥ |M22 | = 443, 520.
We now apply the contraction operation to these groups. Considering the action of M12
on the 12-letter set Ω = {x1 , x2 , · · · , x12 }, we designate some element, say x12 , of Ω as the
distinguished element F in the definition of π 4 . Then define for each π ∈ M12 the permutation
π 4 on the set Ω exactly as in the introduction. Thus, using the natural ordering of elements of
Ω by subscript, the image string of any σ ∈ M12 can be written σ(x1 )σ(x2 ) · · · σ(x11 )σ(F ).
4
As before, we let π− be the permutation on 11 symbols obtained from π 4 by dropping the
final symbol F , and for any subset S ⊂ M12 , we let S−4 = {π− 4
: π ∈ S}, sometimes writing this
4
as (S)− .
Proposition 13 a) hd((M12 )4 − ) ≥ 6.
b) M (11, 6) ≥ |M12 | = 95040.
c) M (10, 6) ≥ 8640.
Proof. We start with a). Suppose not. Since hd(M12 ) = 8, and for any α, β ∈ M12 we have
4
hd(α4 , β 4 ) ≥ hd(α, β) − 3 by Lemma 2a, the contrary assumption implies hd((M12 )− ) = 5.
Thus there is a pair σ, τ ∈ M12 such that hd(σ, τ ) = 8 and hd(σ 4 , τ 4 ) = 5; so hd(σ 4 , τ 4 ) =
hd(σ, τ )−3. Thus by Lemma 2b we know that πσ −1 has a 3-cycle in its disjoint cycle factorization
so the order of πσ −1 is divisible by 3.
Since hd(σ, τ ) = 8 and π and σ are permutations on 12 letters, it follows that there are four
positions, call them xi , 1 ≤ i ≤ 4, at which π and σ agree. Then πσ −1 belongs to the subgroup
H of M12 fixing these four positions; that is H = {α ∈ M12 : α(xi ) = xi , 1 ≤ i ≤ 4}. This H,
denoted M8 , is known to be isomorphic to Q8 , the quaternion group of order 8 ([5], section 3.2).
We can also verify this directly by making use of GAP (Groups, Algorithms, Programming), a
system for computational discrete algebra. The following output employing GAP shows that
H∼ = Q8 , the quaternion group of order 8 ([14])
gap> G := M athieuGroup(12);;
gap> H = Stabilizer(G, [1, 2, 3, 4], OnT uples);;
gap> StructureDescription(H);
“Q008
Now the order of πσ −1 is divisible by 3 as noted above. But 3 does not divide |Q8 |, a
contradiction to Lagrange’s theorem.
Consider next b). Using Lemma 2c and hd(M12 ) = 8 > 3, we have |M12 | = |(M12 )4 − |. Thus
4 4
(M12 )− is a permutation array on 11 letters of size |M12 | with hd((M12 )− ) ≥ 6. Part b) follows.
For part c), we recall from the introduction the elementary bound M (n − 1, d) ≥ M (n,d) n
.
M (11,6)
Using part a), we then obtain M (10, 6) ≥ 11 ≥ 8640.

We remark that using the same method as in part b) of the above proposition one can show
M (10, 6) ≥ |M11 | = 7920. But this is obviously weaker than the bound we give in part c).

16
 We now consider the contraction of M24 and resulting special case bounds for M (n, d). Using
similar notation as for M12 above, we let M24 act on the set of 24 letters Θ = {x1 , x2 , · · · , x24 },
and we designate x24 as the distinguished symbol F in the definition of π 4 from the introduction.
Now define π 4 for any π ∈ M24 as in the introduction, along with accompanying definitions S 4
and S−4 for S ⊆ M24 .

Proposition 14 a) hd((M24 )4 − ) ≥ 14.

b) M (23, 14) ≥ |M24 | = 244, 823, 040.
c) M (22, 14) ≥ |M2324 | = 10, 644, 480.
d) M (21, 14) ≥ |M 24 |
23. 22
= 483, 840.

Proof.
For a), suppose not. Since hd(M24 ) = 16, and for any α, β ∈ M24 we have hd(α4 , β 4 ) ≥
hd(α, β) − 3, it follows that hd((M24 )4− ) = 13. Thus there is pair σ, τ ∈ M24 such that hd(σ, τ ) =
16 and hd(σ , τ ) = 13; so hd(σ , τ 4 ) = hd(σ, τ ) − 3. Hence by Lemma 2b, τ σ −1 has a 3-cycle
4 4 4

in its disjoint cycle structure factorization.

Since hd(σ, τ ) = 16, and σ and τ are permutations on 24 letters, it follows that σ and τ must
agree on 8 positions. Thus τ σ −1 belongs to the subgroup H of M24 fixing these 8 positions.
From the structure theory of M24 , we know that if these 8 positions form an “octad” (among
the 24 positions), then H = M16 ∼ = Z2 × Z2 × Z2 × Z2 , the elementary Abelian group of order 16
([24] Theorem 3.21, and [25] pp. 197-208). Again, this can also be verified directly using GAP
from the following output ([14]).
gap> G := M athieuGroup(24);;
gap> H := Stabilizer(G, [1, 2, 3, 4, 5], OnT uples);;
gap> S = SylowSubgroup(H, 2);;
gap> octad := F iltered([1..24], x → not x in MovedPoints(S));
[1,2,3,4,5,8,11,13]
gap> H := Stabilizer(G, octad, OnT uples);;
gap> StructureDescription(H);
“C2 × C2 × C2 × C200 .
Otherwise, H is the identity ([24], Lemma 3.1). Now the order of τ σ −1 is divisible by 3, so 3
must divide |H|. by Lagrange’s theorem, this contradicts that |H| has order either 16 or 1.
Consider next b). Using Lemma 2c and hd(M24 ) = 16 > 3, we have |M24 | = |(M24 )4 − |. Thus
(M24 )− is a permutation array on 23 letters of size |M24 |, and by part a) we have hd((M24 )4
4
−) ≥
14. Part b) follows.
For part c), we again use the bound M (n − 1, d) ≥ M (n,d) n
. Using part b), we then obtain
M (23,14) |M24 |
M (22, 14) ≥ 23 ≥ 23 = 10, 644, 480.
For d), using M (n − 1, d) ≥ M (n,d)
n
again we get M (21, 14) ≥ M (22,14)
22
≥ |M 24 |
23. 22
= 483, 840.

5 Concluding Remarks
We mention some problems left open from our work.
1. Recall that if I is an independent set in CP (q), then M (q, q − 3) ≥ |I|. To find a large such
I, one can focus on any nontrivial connected component, say P1 , of CP (q). If P1 contains an

17
independent set of size k, then by the isomorphism of the components, we get an independent set
of size k(q − 1) + q(q − 1) = (q − 1)(k + q) in CP (q), where q(q − 1) counts the number of isolated
points in CP (q). Our lower bound M (q, q − 3) ≥ Kq 2 log q implies, again by the isomorphism
of components, that α(P1 ) ≥ Cq log q (where α(G) is the maximum size of an independent set
in a graph G), for some constant C. We therefore ask whether an improvement on this lower
bound for α(P1 ) can be found.
Now V (P1 ) can be viewed as a rectangular array {(i, a) : i, a ∈ GF (q)} as in Figure 1, where
we let i be the row index, and a the column index. By Corollary 8a an independent set in P1 is
just a subset S of this array with the property that for any two points (i, a), (j, b) ∈ S we have
(b − a)(j − i) 6= 1 in GF (q). Using the integer programming package GUROBI, we computed
independent sets in P1 of size k for various q. This k, together with the resulting lower bound
(q − 1)(k + q) for M (q, q − 3) are presented in Table 1.
2. We also ask for good upper bounds on α(P1 ).

q k M (q, q − 3)
7 13 120
13 33 552
19 81 1,800
31 122 4,590
37 191 8, 208
43 191 9,828
49 226 13,200
61 314 22,500
67 340 26,862
73 382 32,760
79 415 38, 532
97 535 60, 672
121 2613 328, 080
127 768 112,770
151 945 164,400
157 984 177, 996
163 1031 193, 428
169 1069 207,984
181 1174 243,900
193 1262 279,360
199 1310 298,782
211 1403 338,940
223 1496 381,618
277 1956 616,308
289 2045 672,192
307 2197 766,224
331 2396 899,910
337 2462 940,464
343 2501 972,648

Table 1: Independent set size k in P1 obtained by integer programming, and resulting lower
bound (q − 1)(k + q) for M (q, q − 3).

18
6 Appendix - Some facts from Number Theory
In this section we review some facts from number theory that were used in this paper.
We start with some notation. For an odd prime p and integer r 6≡ 0(mod p), define the
Legendre symbol ( pr ) to be 1 (resp. -1) if r is a quadratic residue (resp. nonresidue); that is a
square (resp. nonsquare) mod p. If r ≡ 0(mod p), then define ( pr ) = 0. A couple of simple facts
about this symbol are these.

Lemma 15 For an odd prime p and integers r and s we have the following.
a) ( −1
p
) = 1 if p ≡ 1(mod 4), and ( −1
p
) = −1 if p ≡ 3(mod 4).
b) ( rs
p
) = ( r
)(
p p
s
).

Proof. For a), suppose p ≡ 1(mod 4). So write p = 4k + 1, and consider the multiplicative
group of nonzero elements mod p, which has order 4k and is cyclic. Let x be a generator of this
group. Then note that in this group we have 1 = x4k = (x2k )2 , while also (−1)2 = 1 in this
group. Since the quadratic z 2 − 1 = 0 has exactly two solutions z = 1 or −1 in GF (q), and
since x2k 6= 1 since x is a generator, it follows that x2k = −1. Thus -1 is a square mod p.
If p ≡ 3(mod 4), then this cyclic group has order 4k + 2 for some integer k. This time we
have 1 = (x2k+1 )2 , so that by the same reasoning as above we have x2k+1 = −1. This shows
that -1 is not a square mod p, since it is on odd power of the generator.
Consider now b). Just observe that the product rs is a square mod p if and only if both r and
s are squares mod p or if both r and s are non-squares mod p. Part b) then follows immediately.

We now recall the quadratic reciprocity law.

Theorem 16 (Quadratic Reciprocity Law, due to Gauss) For odd primes p and q we have
p q p−1 q−1
( )( ) = (−1)( 2 )( 2 ) .
q p

There are lots of proof of quadratic reciprocity in the literature, so we omit the proof here.
Now let’s apply these facts to determining ( −3
p
) for odd primes p.

Theorem 17 Let p > 3 be an odd prime. Then

a) If p ≡ 1 (mod 6), then -3 is a quadratic residue mod p.
b) If p ≡ 5 (mod 6), then -3 is a quadratic nonresidue mod p.

Proof. By the lemma above we have ( −3

p
) = ( −1
p
)( p3 ), while by quadratic reciprocity we have
p−1
( p3 ) = ( p3 )(−1) 2 . Thus
−3 p−1 −1 p
( ) = (−1) 2 ( )( ).
p p 3
The factors on the right depend on the residue classes of p mod 4 and p mod 3. Thus we
consider the four cases defined by the combinations of these two possibilities, obtaining results
that initially depend on the residue class of p mod 12.
case 1: p ≡ 1(mod 4) and p ≡ 1(mod 3); equivalently p ≡ 1(mod 12).
p−1
Now p ≡ 1(mod 3) says that ( p3 ) = 1. Also p ≡ 1(mod 4) implies (−1) 2 = 1 and by Lemma
15 also implies ( −1
p
) = 1. So by the formula above we have ( −3p
) = 1, showing that −3 is a
quadratic residue when p ≡ 1(mod 12).

19
case 2: p ≡ 1(mod 4) and p ≡ 2(mod 3); equivalently p ≡ 5(mod 12).
p−1
Now p ≡ 2(mod 3) says that ( p3 ) = −1. Also p ≡ 1(mod 4) implies (−1) 2 = 1 and also
Lemma 15 implies ( −1 p
) = 1. So by the formula above we have ( −3p
) = −1, showing that −3 is
a quadratic nonresidue when p ≡ 5(mod 12).
case 3: p ≡ 3(mod 4) and p ≡ 1(mod 3); equivalently p ≡ 7(mod 12).
p−1
Since p ≡ 1(mod 3) we have ( p3 ) = 1. Also p ≡ 3(mod 4) implies (−1) 2 = −1, and also
Lemma 15 implies ( −1 p
) = −1. So by the formula above we have ( −3 p
) = 1, showing that −3 is
a quadratic residue when p ≡ 7(mod 12).
case 4: p ≡ 3(mod 4) and p ≡ 2(mod 3); equivalently p ≡ 11(mod 12).
p−1
Since p ≡ 2(mod 3) we have ( p3 ) = −1. Again p ≡ 3(mod 4) implies that (−1) 2 = −1, and
also that ( −1
p
) = −1. So by the formula above we get ( −3
p
) = −1, showing that −3 is a quadratic
nonresidue when p ≡ 11(mod 12).
Putting together cases 1 and 3, we see that −3 is a quadratic residue mod p when p ≡ 1(mod
6), while cases 2 and 4 show that −3 is a quadratic nonresidue mod p when p ≡ 5(mod 6), as
required.

Corollary 18 Consider the prime power q = pm , where p > 3 is an odd prime. If q ≡ 1(mod
3), then −3 is a square in the finite field GF (q).

Proof. Since p > 3 is an odd prime we have either p ≡ 1(mod 6) or p ≡ 5(mod 6). If p ≡ 1(mod
6), then −3 is already a square in the prime subfield GF (p) ⊆ GF (q) by Theorem 17, so −3 is
a square in GF (q), as required. √
So suppose p ≡ 5(mod 6). Consider the quadratic extension GF (p)( −3) of GF (p) obtained
2
by adjoining to√ GF (p) a root2 of the irreducible (by Theorem2 17) polynomial x + 3 over GF (p).
Then GF (p)( −3) ∼ = GF (p ), and −3 is a square in GF (p ).
Since q ≡ 1(mod 3), then since p ≡ 5(mod 6) we have p ≡ 2(mod 3), so it follows that m
must be even. We recall the basic fact from finite fields that GF (pr ) ⊆ GF (ps ) if and only if
r|s. It follows that GF (p2 ) ⊆ GF (q). Thus since −3 is a square in GF (p2 ), then −3 is a square
in GF (q).

Corollary 19 Let q = pm be a prime power, q ≡ 1(mod 3).

a)The equation x2 + x + 1 = 0 has two distinct solutions in GF (q).
b)For q odd and distinct i, j ∈ GF (q), the equation x2 − (i + j)x + ij + (i − j)2 = 0 has two
distinct roots in GF (q).

Proof. Consider a), and suppose first that p is odd. Since the characteristic of the field
is odd, we may
√ find the solutions
√ by the standard quadratic formula.
√ We obtain the solutions
x = 21 [−1+ −3 ], 12 [−1− −3 ], where we have used the existence of −3 in GF (q) by Corollary
18. These solutions are distinct since p is odd. Pm−1 2i
Now suppose p = 2. Recall the trace function T rGF (q)/GF (2) (x) = i=0 x , defined for
any x ∈ GF (q), which we abbreviate by T r(x). It can be shown (see [21]) that the quadratic
equation ax2 + bx + c = 0, with a, b, c ∈ GF (2m ), a 6= 0, has two distinct solutions in GF (2m ) if
and only if b 6= 0 and T r( ac
b2
) = 0. In our case we have a = b = c = 1, so ac
b2
= 1. Since p = 2 and
q ≡ 1(mod 3), m must be even. Thus there are an even number of terms in the sum defining
T r(x), each of them equal to 1. So since the characteristic is 2, we get T r( ac
b2
) = 0 in our case.
It follows that x2 + x + 1 = 0 has two distinct solutions when p = 2, as required.

20
 Next consider b). Applying p the quadratic formula in this field of oddpcharacteristic, we get
the two solutions x = 21 [ i + j ± (i + j)2 − 4(ij + (j − i)2 ) ] = 21 [ i + j ± −3(i2 + j 2 ) + 6ij ] =
1
p √
2
[ i + j ± −3(i − j)2 ] = 21 [ i + j ± −3(i − j) ]. Now since −3 is a square in GF (q) for
q ≡ 1(mod√ 3) by Corollary
√ 18, it follows that the
√ two solutions√ for x can be written as x1 =
1 1
2
[i(1 + −3) + j(1 − −3)], and x2 = 2 [i(1 − −3) + j(1 + −3)]. Also these two solutions
are distinct since i 6= j and q is odd.

References
[1] N. Alon: Independence numbers of locally sparse graphs and a Ramsey type result, Random
Structures and Algorithms 9(3) (1996), 271-278.

[2] S. Bereg, A. Levy, and I.H. Sudborough: Constructing Permutation Arrays from Groups,
Designs, Codes, and Cryptography 86(5) (2018), 1095-1111.

[3] S. Bereg, L. Morales, and I.H. Sudborough: Extending permutation arrays: improving
MOLS bounds, Designs, Codes, and Cryptography 83(3) (2017), 661-883.

[4] I. Blake, G. Cohen, and M. Deza: Coding with permutations, Information and Control 43
(1979), 1-19.

[5] L.S. Boya: Introduction to Sporadic Groups, Proceedings of the Workshop “Supersymmet-
ric Quantum Mechanics and Spectral Design”, SIGMA 7 (2011), 1-18.

[6] P.J. Cameron: Permutation Groups, Cambridge University Press, Vol. 45 (1999).

[7] C.J. Colbourn, T. Klove, Alan C.H. Ling: Permutation arrays for powerline communication
and mutually orthogonal latin triangles, IEEE Transactions on Information Theory 50(6),
1289-1291.

[8] W. Chu, C.J. Colbourn, and P. Dukes: Constructions for permutation codes in powerline
communications, Designs, Codes, and Cryptography 32 (2004), 51-64.

[9] M. Deza and S.A. Vanstone: Bounds for permutation arrays, J. of Statistical Planning and
Inference 2(2) (1978), 197-209.

[10] J. Dixon and B. Mortimer: Permutation Groups, Graduate Texts in Mathematics Vol.
163, (1996) Springer-Verlag, New York.

[11] P. Dukes and N. Sawchuk: Bounds on permutation codes of distance four, J. of Algebr.
Comb. 31 (2010), 143-158.

[12] H.C. Ferreira and A.J.H. Vinck: Interference cancellation with permutation trellis arrays,
Proc. IEEE Vehicular Technology Conf., Boston, MA, USA (Sept. 2000), 2401-2407.

[13] P. Frankl and M. Deza: On the maximum number of permutations with given maximal or
minimal distance, J. of Combinatorial Theory, Series A 22(3) (1977), 352-360.

[14] D.F. Holt, personal communication.

21
[15] S. Huczynska: Powerline communications and the 36 officers problem, Philosophical Trans-
actions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences
364(1849) (2003), 34-40.

[16] I. Janiszczak, W. Lempkin, P.R. Ostergard, and R. Staszewski: Permutation codes invariant
under isometries, Designs, Codes, and Cryptography 75(3) 2015, 497-507.

[17] A. Jiang, R. Mateescu, M. Schwartz, and J. Bruck: Rank modulation for flash memories,
Proc. IEEE Symp. Inf. Theory (2008), 1731-1735.

[18] A. Jiang, M. Schwartz, and J. Bruck: Error-correcting codes for rank modulation, Proc.
IEEE Symp. Inf. Theory (2008), 1736-1740.

[19] P. Keevash and C.Y. Ku: A random construction for permutation codes and the covering
radius, Designs, Codes, and Cryptography 41 (2006), 79-86.

[20] N. Pavlidou, A.J.H. Vinck, J. Yazdani, and B. Honary: Powerline communications: state
of the art and future trends, IEEE Commun. Mag. 41(4) (2003), 34-40.

[21] K. Pommerening: Quadratic equations in finite fields of characteristic 2, unpublished

manuscript (2000), English version (February 2012).

[22] D.H. Smith and R. Montemanni: A new table of permutation codes, Designs, Codes, and
Cryptography 63(2) (2012), 241-253.

[23] D.R. Stinson: Combinatorial Designs, Springer, India (2010).

[24] Leo Taslaman, The Mathieu groups, M.S. thesis, Lund University, 2009.

[25] T. Thompson: From error-corecting codes through sphere pacing to simple

groups, Carus Mathematical Monographs 21, Mathematical Association of America (1983).

[26] D.R. de la Torre, C.J. Colbourn, and A.C.H. Ling: An application of permutation arrays
to to block ciphers, Proc. 31st Southeastern Int. Conf. on Combinatorics, Graph Theory,
and Computing, Boca Raton, Fl., USA 145 (2000), 5-7.

[27] A.J.H. Vinck: Coded modulation for powerline commumnications, A.E.U. Int. J. Electronic
Commun. 54 (2000), 45-49.

[28] X. Wang, Y. Zhang, Y. Yang, and G. Ge: New bounds of permutation codes under Ham-
ming metric and Kendall’s τ -metric, Designs, Codes, and Cryptography 85(3) (2017), 533-
545.

[29] https://en.wikipedia.org/wiki/Mathieu group M24

22